Transcript Slide 1
Nonlinear Systems: Properties and Tests M. Sami Fadali Professor EBME University of Nevada Reno 1 Outline Linear versus nonlinear. Nonlinear behavior. Controllability and observability. Stability. Passivity. 2 State Variables Minimal set of variables that completely describe the system. State: set of numbers (initial conditions) that allows us to solve for the response for a given input. State variable: variables obtained by letting the state evolve with time. Example: position, velocity. 3 Linear State-space Model State equations: Set of linear first-order differential equations. Output equations: Set of algebraic equation. dx1 x2 dt x f x, u, t dx2 2 2 x x x y h x, u, t 1 2 2u dt y x1 u 4 Linear State-space Model Linear equations. Can be solved analytically. x Ax Bu y Cx Du 1 x1 0 x1 0 x 8 6 x 1u 2 2 y 1 1x 5 Linear Systems Additivity: add responses for added effects. Homogeneity: scale responses for scaled effects. Zero-input Response: Due to initial conditions. Zero-state Response: Due to the input. Total response = zero-input response + zero-state response At t 0 x(t ) e x(t0 ) e At Bu( )d t0 zero input response t zero state response 6 Additivity & Homogeneity Response to Initial Conditions Step Response 3 0.5 0.45 2.5 0.4 2 0.35 0.3 Amplitude Amplitude 1.5 1 0.25 0.2 0.5 0.15 0 0.1 -0.5 -1 0.05 0 0.5 1 1.5 2 Time (sec) Zero-input response. 2.5 3 0 0 0.5 1 1.5 2 2.5 3 Time (sec) Zero-state response. 7 Nonlinear Systems x f x, u, t y hx, u, t x f x, t Gx, t u y hx, t Dx, t u No additivity or homogeneity. Dependent responses due to initial conditions and input. More complex behavior 8 Examples of Nonlinear Behavior Multiple equilibrium points. Limit cycles: fixed period without external input. Bifurcation: drastic changes of behavior with small changes in parameter values. Chaos: aperiodic deterministic behavior which is very sensitive to its initial conditions. Response to sinusoid: harmonics, subharmonics or unrelated frequencies. 9 Multiple Equilibrium Points Equilibrium: stay there if you start there. Stability of equilibrium not system. No change: time derivative is zero. Solve for equilibrium points. x 0 f x,u, t 10 Examples Pendulum: Two equilibrium points. Bistable Switch 3 equilibrium points (0, v0, v1) dv g (v ) dt g (v) vv v0 v v1 g(v) v 11 Limit Cycles Unlike linear system oscillations 2 1.8 1.6 1.4 x2 Amplitude does not depend on initial state. Stable or unstable limit cycle. Solution of limit cycle 2.2 1.2 1 dx1 1 x1 x2 dt dx2 x1 1x2 dt 0.8 0.6 0.4 0.2 0.2 0.4 0.6 0.8 1 1.2 x1 1.4 1.6 1.8 2 12 2.2 Example: Fitzhugh-Nagumo Model Simplified version Param v0 v1 of H-H model. 0.2 1 Value Parameters I k b 1 0.5 0 1.05 1.04 dv g (v ) w I dt dw 1 v kw b dt g (v) vv v0 v v1 1.03 1.02 1.01 1 0.99 0.98 0.97 0.96 0.95 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 13 Bifurcation Bifurcation point: Behavior d 2 x x x3 0 dt2 changes drastically as 3 parameter changes slightly. xe xe 0 xe 0, Example: As parameter xe changes: periodic oscillations Stable period doubling chaos. Example: Pitchfork Undamped Duffing Equation Stable Unstable Stable 14 Chaos Behavior is extremely sensitive to initial conditions. Behavior is deterministic but looks random. Example: cardiac arythmia (irregular beating patterns) 15 Lorenz attractor Two unstable equilibrium points. Model turbulent convection in fluids (weather patterns). 16 Response to Sinusoid Linear: scale amplitude and phase shift. Nonlinear: Harmonics: multiple of input frequency. Subharmonics: fraction of input frequency. Unrelated frequency. Examples sin 2 x 1 cos2 x 2 sin x 1 cos2 x 2 17 Response to Noise Linear Systems Gaussian input gives Gaussian output. Completely characterized by mean and covariance matrix (variance). Total response = zero-input response + zero-state response Nonlinear systems Gaussian input gives non-Gaussian output. Need higher order statistics. 18 Example: Chi-Square Distribution 2 x m x 1 f X ( x) 2 1/ 2 exp 2 2 x 2 x x t 2 2 1 2 ( x) e fX(x) dt x 0.2 n y xi2 , xi ~ N 0,1 i 1 fY(y) 0.18 0.16 0.14 fY ( y) 2 0, 1 n/2 ( n / 2) (u ) t 0 y ( n / 2 ) 1e y / 2 , y 0 y0 n=4 D.O.F. 0.12 0.1 0.08 0.06 e dt u 1!, u integer u 1 t 0.04 y 0.02 0 0 20 40 60 80 100 120 140 160 180 200 19 System Properties Stability Controllability Observability Passivity 20 Robustness Property holds over a specified subset of parameter space. Sensitivity: local measure of robustness. Robustness w.r.t. noise and disturbances. 21 Bode Sensitivity per unit changein F F F S Lim Lim p 0 per unit changein p p 0 p p p F Lim F p 0 p F p p F p F F p F p small p F F F p S p , small p p 22 Example:Biochemical System Metabolite Xi is produced from substrate Xj by an enzyme-catalyzed reaction (MM Kinetics) V X max j X i v Km X j S v Xj Km X j X j Km X j v X j Vmax 2 X j v Vmax K m X j Km Km X j 23 Sensitivity Equation First-order estimates of the effect of parameter variations (near q*) x f x, t , q S : x q S At , q S Bt , q , S (t 0 ) 0 f At , q x q f Bt , q q q 24 Stability Local or global Lyapunov stability: continuity w.r.t. the initial conditions. Asymptotic stability: Lyapunov stability plus asymptotic convergence to the equilibrium. Exponential stability: ||x|| trajectory bounded above by an exponential decay. 25 Stability 2 1.8 Exponential 1.6 Unstable 1.4 Stability 1.2 1 0.8 0.6 0.4 Asympt.Stable 0.2 0 0 0.5 1 1.5 2 2.5 3 Stable i.s. L. 26 Example: Model of Linear Pathway Specify kinetic orders, independent variables Determine equilibrium: (1/4, 1/16, 1/64) Solve differential equations (separation of variables): asymptotically stable. X4 X 1 0.5 X 4 X 10.5 X 0.5 X 0.5 4 X 2 1 X 3 4 X 2 2 X X 4 0 .5 X1 2 0.5 3 X2 X3 Equilibrium: (0, 0, 0) x1 X 1 1 4 x2 X 2 1 16 x3 X 3 1 64 0.5 x1 0.5x1 1 4 0.5 x2 0.5x1 1 4 4x2 1 16 0.5 x3 4x2 1 16 2x3 1 64 27 Stability of Motion Stability of equilibrium of the error dynamics 2 1.5 1 x f x, t x m f m x m , t Error 0.5 0 e x xm e f x m e, t f m x m , t f e e, t -0.5 0 0.5 1 1.5 2 2.5 3 28 Lyapunov Stability Theory Generalized energy function (positive definite). Energy min at a stable equilibrium, energy max at an unstable equilibrium. Trajectories converge to equilibrium if energy is decreasing in its vicinity (negative definite). Design: choose control to make energy decreasing along trajectories. 29 Laypunov Stability Theorem Asymptotic stability if x f (x) V ( x) 0 T V (x) V x f 0 x(k 1) f (x(k )) V (x(k )) 0 V (x(k )) V (x(k 1)) V (x(k )) 0 30 Lyapunov Approach Use quadratic Lyapunov function. Local stability for v < 0.2 Negative definite f(v) derivative 1 2 v 40w2 2 dV dv dw v w dt dt dt v 4 1.2v 3 0.2v 2 vw wv 0.5w2 V v dv g (v ) w dt dw 1 v w 2 dt 40 g ( v ) v 3 1 .2 v 2 0 .2 v 0 v 4 1.2v 3 0.2v 2 0,0,1,0.2 31 Method to Obtain a Lyap Function Krasovskii’s method: Use Jacobian (derivative) of RHS of state eqn. Stable if the derivative is negative near the origin. x f (x) V (x) f T Pf 0 V (x) f T AT Pf f T PAf f T Ff 0 F AT P PA 0 A f x 32 Example: Metabolic Process Use Jacobian (derivative) of RHS of state eqn. 101 1 0 P 1 1 0 0 0 1 1 2 4 x1 1 1 A 2 4 x1 1 0 0 0 4 0 4 4 64x3 1 0.5 x1 0.5x1 1 4 0.5 x2 0.5x1 1 4 4x2 1 16 0.5 x3 4x2 1 16 2x3 1 64 100 4 x1 1 F 4 0 8 4 0 16 4 64x3 1 4 0 33 Example: Fitzhugh-Nagumo Model System with zero bias has a stable equilibrium (stable node) at (0,0). Small perturbation: return to equilibrium. 1.5 dv g (v ) w dt dw 1 v w 2 dt 40 g (v) vv 0.2 v 1 1 0.5 0 -1 -0.5 0 0.5 1 1.5 -0.5 0 0.5 1 1.5 -0.5 -1 -1.5 -1 34 Limitations Sufficient conditions for stability and instability: if condition fails, no conclusion. Necessary and sufficient for the linear case only. A P PA Q T A PA P Q T Q 0, P 0 35 Controllability & Observability Controllability: Can go wherever you want no matter where you start. x0, xf , control u:[0,T]U, T < , s.t. x(T; x0) = xf. Indistinguishable: u U x01, x02, y:[0,T]Y, T < y(T, x01) = y(T, x02) Observability: Can determine the initial state from the measurements (no two are indistinguishable). x01, x02, y(T, x01) = y(T, x02) x01= x02. 36 Graphical Interpretation x2 Uncontrollable Subspace Controllable Subspace u x2 Unobservable Subspace x1 Observable Subspace y x1 37 Example Identical tanks with identical connections to a water source. Not observable: Measuring the difference gives zero regardless the two levels. Not controllable.: Filling the two tanks from one source gives the same level. 38 Passivity Supply rate: integrate to obtain energy. Storage function: S Dissipative system: storage < supply Passive: dissipative with bilinear supply rate. S 0 0 S x(t ) 0 S x(T ) S x(0) uT ydt T 0 T S u y 39 Example of Passive System Spring-massdamper R-L-C circuit. mv bv k x f vbv 0 k x 0 1 2 S m v k x dx 2 dS dS dx dS dv dt dx dt dv dt k x v m vv v f k x bv k x v vf vbv 0 40 Zero Dynamics Internal dynamics of the system when the output is kept identically zero by the input. Example: Metabolite Concentrations Select X4 such that X1 = 0 how do X2 & X3 behave? X 1 0.8 X 21 X 31 X 40.5 3 X 10.5 X 20.1 2 X 10.75 X 30.2 X 3 X 0.1 X 0.1 2 X 0.5 2 1 2 2 X 3 2 X 10.75 X 30.2 5 X 30.5 X4 0 X 2 2 X 20.5 X 5 X 0.5 3 3 41 Stability of Passive Systems Zero-state detectable (observable) System with zero input has stable zero dynamics (resp. y=0 x=0) Theorem: Zero-state detectable and passive a) x=0 with u=0 is stable. b) x=0 with u= y= h(x) is asymptotically stable. 42 Absolute Stability (e) Stable for any sector-bound nonlinearity. G linear passive e e u G (.) y 43 Example: Artificial Neural Networks Use passivity to show stability V 1 b e Z-1 y p 44 Passivity of Linear Systems (CT) A minimal state-space realization (A, B, C, D) is passive if and only if there exist real matrices P, L, and W such that PP 0 T AT P PA LT L PB C L W T T W W DD T T 45 Passivity of Linear Systems (DT) A minimal state-space realization (A, B, C, D) is passive if and only if there exist real matrices P, L, and W such that PP 0 T AT PA P LT L AT PB C T LT W W W D D B PB T T T 46 Passivity of Periodic System (F, G, H, E) =DT minimal cyclic reformulation of a periodic system. System is passive if and only if it satisfies the following conditions with a positive definite symmetric matrix P real matrices W and L. 47 Periodic KYP Pi AT (i) Pi 1 A(i) Qi , i 0,, T 1, PT P0 P0 T (T ,0) P0(T ,0) Qs T 1 Qs T (i,0)Qi (i,0) i 0 LTi Wi C T (i) AT (i) Pi 1 B(i), i 0,, T 1 WiT Wi D(i) DT (i) BT (i) Pi 1 B(i), i 0,, T 1 48 Linearization Local behavior in the vicinity of an equilibrium. Stability. Controllability. Observability. Passivity: KYP lemma. 49 Linearization 1st order approximation f(x) df f ( x) f ( x0 ) x O(2 x) dx x x0 x x x0 f m x, for small x x f x f x f x i x x x x x0 x j x 0 80 f(x0) 60 40 20 0 0 2 4 6 x0 8 x 50 Linearization of Linear Pathway Equilibrium: (1/4, 1/16, 1/64) x f x f x x x Ax x x0 X 1 0.5 X 4 X 10.5 X 0.5 X 0.5 4 X 2 1 X 3 4 X 2 2 X 30.5 X 4 0 .5 2 X 1 1 2 0 0 X 1 X X 1 2 4 0 2 2 X 3 0 4 1 8 X 3 Stable Equilibrium: (1/2, 4, 1/8) all in LHP 51 Stability Stability Condition: Eigenvalues in LHP dx1 x2 dt dx2 x1 1 x2 1 dt Equilib. x1 1 x2 0 x f x f x x x Ax x x0 1 z1 z1 0 z 1 1 z 2 2 2 1 0 52 Stability of Linear DT Systems Eigenvalues inside the unit circle. Examples x(k 1) x(k ) stable, a 1 x(k 1) ax(k ) unstable, a 1 Im[z] Unit Circle STABLE Re[z] UNSTABLE 0 0 x1 (k ) x1 (k 1) 1 2 x (k 1) 1 2 0.4 x (k ) 0 2 2 x3 (k 1) 0 4 1 8 x3 (k ) 53 Conditions for 2nd-order Case Second-order recursion a2 x(k 2) a1 x(k 1) a0 x(k ) 0, a2 0 a2 a1 a0 0 a2 a0 0 a2 a1 a0 0 a0 1 stable 1 a1 1 1 54 Example: Dynamic Neural Network IIR Filter Nonlinear activation function (monotone increasing, slope =g2 >0) Stable network for any stable matrix A. Problem: How to minimize error subject to the stability constraint? bi0 xi(k+1) u wi + z1 xi(k) y bi + f (.) A 55 Constrained Optimization Minimize square error subject to stability constraints. Consider 2nd-order (explicit constraints) Modify: stability margins (safety factor) 2 J min d i ai a0 , a1 s.t. i 1 2 1 a1 a0 0 1 a0 0 1 a1 a0 0 56 Global Linearization Find a transformation of the nonlinear system to a decoupled linear system (easy transformation is special cases). Design linear control then transform back. Use differential geometry to derive the theory. 57 Example: Mechanical Systems D(q)q C (q, q )q g(q) q = vector of generalized coordinates. D(q) = ss positive definite inertia matrix C (q, q ) = ss matrix of velocity related terms g(q) = s1 vector of gravitatioinal terms = vector of generalized forces 58 Global Linearization Let the acceleration vector be the input. Series of double integrators. Choose acceleration for desired behavior. Calculate torque from accelerations, positions, and velocities. (t ) u(t ) q D(q)u(t ) C (q, q )q g(q) 59 Limitations Complex mathematical theory (general case) but solution is hard to obtain: must solve a nonlinear partial differential equation analytically for transformation. Results sensitive to modeling errors. Nonlinearity and coupling can be exploited to provide desirable behavior. 60 Discrete-time Periodic Systems x(k 1) Ad k x(k ) Bd k uk y(k ) C (k ) x(k ) D(k )u(k ) All system matrices are governed by M(k) = M(k+K) , k = 0, 1, 2, ... Model multi-rate sampled systems. Time-invariant reformulations: lift, cyclic. 61 Fuzzy Models Include qualitative information. Fuzzy sets: graded membership. Ai Ai+1 … … x ei1 ei ei+1 62 Lyapunov Stability of TS Systems Linear matrix inequalities (LMI). Common Lyapunov function. Restrictive: system can be stable even if one or more local model is unstable! Computational load: large number of LMIs. A P PA 0 T A PA P 0 T Q 0 63 Hybrid Systems Switch between different models: Includes piecewise-linear. Overall behavior can be: Stable even if each subsystem is unstable. Unstable even if each subsystem is stable. Piecewise linear: Sufficient stability condition using common Lyapunov function. 64 Example:Gene Regulation Effector gene cycles between two alternative environments: H = high demand environment (negative control mode: repressor protein). L = low demand environment Positive control mode: activator protein). TH (TL)Av. duration in phase H (L). Av. Cycle time = C = TH + TL 65 Mathematical Model AH x, x H x AL x, x L x(t 0 TH ) exp(AH TH ) x(t 0 ) x(t 0 TH TL ) exp(ALTL ) exp(AH TH ) x(t 0 ) x(t0 lC) AHL x(t0 ) l Response diverges if an eigenvalue of AHL is greater than 1. Steady state: zero if all eigenvalues are inside the unit circle. Nonzero for one or more eigenvalue on the unit circle. 66 Simulation Example 0 1 1 0 AH A L 1 0.1 1 1 1 x1 (k ) x1 (k 1) 0 x (k 1) 0.1 1.1 x (k ) 2 2 1 1, 2 0.1 1 v1 1 x Lss AL x Hss 1 v2 0.1 1 0 1 1 1 0.1 1 1.1 67 Fault Detection Predict output of system using a mathematical model. Compare predicted output to measured output: primary residual. Filter the primary residual and use the result to detect an error. Use state estimator (Kalman filter, observer, Bayesian network, fuzzy model, neural network) 68 Conclusion Mathematical models of physical systems Linear. Nonlinear. Piecewise-linear. Properties Stability. Controllability & observability. Passivity. Applications. 69 70